Unverified50% confidenceBenchmarkExact time
At 1M length, RecSA scores 60.5 on the RULER benchmark, while the best baseline Dense + DCA scores only 40.0.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
11/6/2026
First Seen
Valid until: 11/20/2026
Sources
Related Entities
Related Claims
UnverifiedWith optimized kernels, RecSA's end-to-end prefill and decoding from 64K to 2M are both faster than dense attention.66% similarUnverifiedOn a non-overlapping validation slice, Continuum Attention achieves a 0.0457 nats/byte advantage over the strongest softmax baseline (a screened version), accompanied by a paired 95% confidence interval.62% similarUnverifiedMLA(Multi-head Latent Attention)通过低秩压缩将 KV 缓存的显存占用降低至传统 MHA 的约 1/859% similarUnverifiedWhen extending the context from 192,000 to 256,000, the expert cache dropped from 16.44 GB to 13.66 GB and KV increased from 4.31 GB to 7 GB, with almost zero speed cost, because KV is fully in VRAM and the expert cache automatically balances.58% similarVerifiedTaking the LLaMA-2 70B model as an example, a sequence of 2048 tokens requires approximately 4 GB of VRAM for the KV cache alone.58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1141965API
curl https://kongchang.com/api/v1/knowledge/claims/1141965MCP
get_claim(id=1141965)